AggregationNode's memory estimate should be based on NDV only for non-grouping aggs

Currently, the AggregationNode always computes the NDV to estimate the number of rows. However, for grouping aggregates, the entire input has to be consumed before the output can be produced, hence its memory estimate should not consider the NDV. This is acceptable for non-grouping aggregates because it only need to store the value expression during the build phase, instead of the entire tuple.